Real-time Pose Estimation in Mobile with Dense Upsampling Convolution

Yingxian Chen, Baoheng Zhang, Wilton W. T. Fok

Abstract: Human pose estimation (HPE) has been gradually applied to our daily life. It’s significant to design a simple yet effective model structure for real-time HPE. Several backbones are available for pose estimation, but many of them are imprecise, complex, and mislocalized when it comes to reconstruction. In order to narrow the gaps, several recent studies have demonstrated that deconvolution reconstruction is highly effective in achieving high levels of accuracy. Using the current popular backbones, we re-analyze and reconstruct the models. The efficiency and accuracy are state-of-the-art. Additionally, we release a new dataset that represents real-world data related to yoga. As a result of the development of our framework, we are able to achieve improvements in our released Yoga dataset named SAILPOSE-YOGA as well as other existing benchmarks for the estimation of single poses. The dataset will be released on https://github.com/carolchenyx/SAILPOSE-YOGA.git